Computers in Biology and Medicine 88 (2017) 50–59

Contents lists available at ScienceDirect

Computers in Biology and Medicine

journal homepage: www.elsevier .com/locate/compbiomed

Automatic discrimination of actinic keratoses from clinical photographs

Panagiota Spyridonos a,*, Georgios Gaitanis b, Aristidis Likas c, Ioannis D. Bassukas b

a Department of Medical Physics, Faculty of Medicine, School of Health Sciences, University of Ioannina, University Campus, 45110 Ioannina, Greeceb Department of Skin and Venereal Diseases, Faculty of Medicine, School of Health Sciences, University of Ioannina, University Campus, 45110 Ioannina, Greecec Department of Computer Science & Engineering, University of Ioannina, University Campus, 45110 Ioannina, Greece

A R T I C L E I N F O

Keywords:Actinic keratosisClinical photographyColor texture analysisLocal binary patternsTextonsSupport vector machines

* Corresponding author.E-mail addresses: pspyrid@cc.uoi.gr (P. Spyridonos), g

http://dx.doi.org/10.1016/j.compbiomed.2017.07.001Received 1 May 2017; Received in revised form 29 June

0010-4825/© 2017 Elsevier Ltd. All rights reserved.

A B S T R A C T

Background and Objective: Actinic keratoses (AK) are common premalignant skin lesions that can progress toinvasive skin squamous cell carcinoma (sSCC). The subtle accumulation of multiple AK in aging individuals in-creases the risk of sSCC development, and this underscores the need for efficient treatment and patient follow-up.Our objectives were to develop a method based on color texture analysis of standard clinical photographs for thediscrimination of AK from healthy skin and subsequently to test the developed approach in the quantification offield-directed treatment interventions.Methods: AK and healthy skin in clinical photographs of 22 patients were demarcated by experts and regions ofinterest (ROIs) of 50 � 50 pixels were cropped. The data set comprised 6010 and 13915 ROIs from AK andhealthy skin, respectively. Color texture features were extracted using local binary patterns (LBP) or texton fre-quency histograms and evaluated employing a support vector machine (SVM) classifier. Classifier evaluation wasperformed using a leave-one-patient-out scheme in RGB, YIQ and CIE-Lab color spaces. The best configuration ofthe SVM model was tested using 157 AK and 216 healthy skin rectangular regions of arbitrary size. AK treatmentoutcome was evaluated in an additional group of eight patients with 32 skin lesions.Results: The best configuration of the discrimination model was achieved by employing LBP color texture de-scriptors estimated from the Y and I components of the YIQ color space. The sensitivity and specificity of the SVMmodel were 80.1% and 81.1% at ROI level and 89.8% and 91.7% at region level, respectively. Based on theclassifier results the quantitative AK reduction was 83.6%.Conclusions: It is important that patients with AK seek evaluation for treatment to reduce the risk of diseaseprogression. Efficient patient follow-up and treatment evaluation require cost-effective and easy to use ap-proaches. The proposed SVM discrimination model based on LBP color texture analysis renders clinical photog-raphy a practical, widely available and cost-effective tool for the evaluation of AK burden and treatment efficacy.

1. Introduction

Actinic keratoses (AK) are common premalignant lesions of the skin,mostly affecting individuals of European ancestry [1] and constitute asignificant workload in dermatology outpatient clinics worldwide [2].AK result from chronic exposure to ambient ultraviolet radiation, thusthey are mostly located on the chronically sun-exposed skin of olderadults. With increasing age susceptible individuals develop multipleclinical and subclinical AK lesions that coexist in carcinogen-exposedskin areas (“field cancerization”) [3]. The biological behavior of an in-dividual AK may vary: it can remain relatively stable in form and size fora long period or spontaneously involute and ultimately disappear. On theother hand, some lesions may evolve into a hyperplastic, hyperkeratoticstate and ultimately a minority of them (<0.1% of lesions/per year) may

gaitan@cc.uoi.gr (G. Gaitanis), arly@

2017; Accepted 2 July 2017

progress to skin squamous cell carcinoma (sSCC), a potentially lethaltumor [4,5]. However, since 60%–80% of sSCCs arise in AK fields, timelytreatment of these lesions is anticipated to prevent progression [6–8].

In practice, and during the examination of a skin field, it is commonfor the clinician to document the presence of multiple, unevenlydistributed, partly coalescent AK of different sizes. In these cases, eval-uating disease burden at baseline and quantifying treatment efficacy is achallenging, real-life problem. The latter is highlighted in correspondingclinical studies where the use of subjective means to measure AK burdenresults in suboptimal interobserver agreement [9–11].

Research on computer vision systems for the evaluation of skin dis-eases is a continuously growing subfield of medical image analysis[12–14]. High-quality clinical images and specialized instruments thatmagnify deeper skin tissues have been coupled with automated detection

cs.uoi.gr (A. Likas), ibassuka@cc.uoi.gr (I.D. Bassukas).

P. Spyridonos et al. Computers in Biology and Medicine 88 (2017) 50–59

systems to offer valuable computer-aided diagnostic tools for theassessment of keratinocytic skin malignancies. To date, non-invasiveimaging of AK has addressed the identification and differential diag-nosis of isolated skin lesions, employing techniques such as the cost-effective digital dermoscopy or the more sophisticated reflectanceconfocal microscopy (RCM) and high-definition optical coherence to-mography (OCT). The diagnostic accuracy of OCT images has beenassessed using ensemble classifiers and support vector machine (SVM)models in the differentiation of AK from basal cell carcinomas (BCC)[15]. Recently, the in vivo optical properties of facial AK subtypes andsSCC were quantified by high-definition OCT and a decision tree diag-nostic algorithm for the discrimination of sun-damaged facial skin fromAK subtypes and SCC was proposed [16]. Likewise, classification treeswith morphologic RCM characteristics have been used for the discrimi-nation of AK from normal skin [17] and to combine dermoscopic signsthat predict the relevant histopathological findings in AK diagnosis [18].Also, a Bayesian classifier and 3D features obtained by a sophisticatedstereo image system have been used to automatically distinguish be-tween keratinocytic malignancies [19]. In addition, rather cumbersome,non-portable equipment employing cross-polarized light and fluores-cence has been proposed for skin cancer screening purposes, includingthe diagnosis of AK [20].

Conventional skin biopsy and histopathological examination remainthe gold standard for confirming the diagnosis of AK. In general, most ofthe aforementioned non-invasive imaging approaches intend to substi-tute biopsy for the diagnosis of selected skin lesions [21]. However,regarding AK, the fundamental limitation of all aforementioned methodsis that they are spatially elective, capable of supporting the recognition offew selected lesions per-patient and thus unsuitable for “field” quantifi-cation purposes. For example employing dermoscopy a relative smallskin area can be evaluated per field of view (corresponding to approxi-mately 4 cm2 for a common commercially available digital dermato-scope; DermLite, PhotoSystem, 3Gen, LLC, Dana Point, CA, U.S.A.).Clinical photography on the other hand can provide morphological in-formation of whole anatomic skin regions that in addition can harbormultiple AK, e.g. the skin of the face or the balding scalp, a quasi ‘scan-ning modality’ for morphological alterations of the skin surface.

The applicability of standard clinical photographs in the evaluation ofkeratinocytic premalignancies has been addressed in a limited number ofstudies. Using clinical photographs a hierarchical classification systembased on the k-nearest neighbors (K-NN) model for discrimination be-tween benign and malignant skin growths has been proposed [22].Furthermore, automatic delineation of AK areas on clinical photographshas been elaborated using color space transforms and morphologicalfeatures for erythema detection [23].

In this study, we developed an SVM model for discrimination of AKfrom normal skin, based on color texture analysis of non-standardizedclinical photographs. Our aim was to use photography to quantify AKburden and to evaluate the outcome of treatment interventions thattarget skin cancerization fields. To the best of our knowledge, the presentstudy is the first image analysis approach towards the quantification ofAK burden for efficient treatment evaluation and patient follow-up bymeans of clinical photography.

2. Materials and methods

2.1. Acquisition of clinical photographs

Institutional approval was granted and patients with at least onebiopsy-proven AK were recruited from a dermatology outpatients clinic;all patients gave informed consent for the photographic assessment oftheir skin lesions. A total of 30 patients (24 men) were included (meanage: 78 [range: 68–85] years). Photographs from 22 patients were usedfor the model development and the rest for evaluating AK burdenreduction after treatment. Photographs were acquired using a NikonD610 camera with a spatial resolution of 6016 � 4016 pixels. A 60 mm

51

prime lens was adapted with two adjustable crossed polarized filters tominimize unwanted glare and a SIGMA EM-140 Macro ring flash. Imageswere rescaled to an equal final size of 50 pixels per millimeter employingas internal standards stickers attached to the skin (MACO Round ColorCoding Labels, 0.635 cm [1/4 inch] in diameter, USA).

The same physician took all images at baseline and during follow-upvisits with the volunteers seated. A detailed description of image acqui-sition is given in Fig. 1.

2.2. Discrimination model implementation

The automatic discrimination of facial AK lesions from healthy skinwas addressed as a two-class classification problem, using theSVM classifier.

SVM is an advanced and extensively used classification method thathas been successfully applied to a variety of real-world data analysisproblems (text categorization, image recognition, bioinformatics andmedical decision), mostly providing improved results compared withother techniques [24–26]. SVM is currently considered the standardmethod used to build a classifier from training data, especially in prob-lems with continuous input features, as in our case. Further details on theSVM classifier can be found in Ref. [27].

In this study, the radial basis function (RBF) kernel was used for theSVM implementation. The SVM classifier performance depends on thechoice of the parameter (C) which is also known as box constraint and thescaling factor (γ) which is the inverse width of the RBF kernel. We testedvarious pairs of (C, γ) values and we selected the one with the best pre-dictive accuracy for both classes.

To train and validate the SVM classifier, AK and healthy skin 50 � 50pixel regions of interest (ROIs) were cropped from areas demarcated byexperts in each photograph (Fig. 2). A 50 � 50 pixel ROI is about onemm2 and is the largest ROI adequate to sample the smallest demarcatedby experts skin area. The data set comprised 6010 and 13915 ROIs of AKand healthy skin, respectively, extracted from 22 patients (Table 1).

Two different SVM models were constructed and evaluated. In thefirst model, each ROI is represented by its color local binary pattern (LBP)histogram (SVMLBP model). In the second model, each ROI is representedby its texton frequency histogram (SVMTextons model).

Both SVM models were evaluated in terms of their sensitivity andspecificity. For this, the number of true positives (TP), false negatives(FN), true negatives (TN) and false positives (FP) was obtained by eachmodel. TN is the number of healthy skin ROIs correctly identified, FN isthe number of AK ROIs incorrectly identified as healthy skin, TP is thenumber of AK ROIs correctly identified and FP is the number of healthyskin ROIs incorrectly identified as AK. Sensitivity is the probability thatthe SVM model will respond positively when tested on the AK ROI:

Sensitivity ¼ TPðTPþ FNÞ (1)

Specificity is the probability that the SVM model will respond nega-tively when tested on the healthy skin ROI:

Specificity ¼ TNðTN þ FPÞ (2)

Finally, the accuracy is defined as the probability that the SVMmodelwill correctly classify both classes:

Accuracy ¼ TPþ TNTPþ FPþ TN þ FN

(3)

Evaluation of the discrimination models was performed using theleave-one-patient-out scheme, meaning that ROIs from all patients butone were used for the training and ROIs from the patient excluded fortesting the model. To avoid overtraining the model in favor of the ma-jority class (the healthy skin class), equal numbers of healthy skin and AKpatterns were randomly selected from each patient in the training set.

Fig. 1. For facial pictures, portrait images with the camera axis perpendicular to the photographed region at 0� (a), 45�/�45� (b) and 90�/�90� (c) angles to the coronal (sagittal) planewere acquired. The distance was adjusted to include the whole face from the chin to the top of the hair/scalp. For balding scalps three sets of landscape pictures were taken: one with thelower frame on the eyebrows and with an angle that positioned the ears in the middle of the frame (d). With this approach images from the forehead and the anterior half of the vertex scalpwere suitable for analysis. Two additional images from both sides of the scalp were taken to include the ipsilateral vertex and forehead. For these images, the landscape frame waspositioned on the eyebrows (lower frame) and the ipsilateral ear (the lateral frame) (e). Likewise, the nuchal area and the top of the scalp were photographed from the back, bothperpendicularly to the nuchal area and at an angle that would position the lower frame on the hairline and the ears in the middle of the frame (f).

P. Spyridonos et al. Computers in Biology and Medicine 88 (2017) 50–59

The best configuration of the SVM model was also tested at skin re-gion level. That is, from each patient, up to 10 arbitrarily sized (box)regions of AK and healthy skin were selected. In total, 157 AK and 216healthy skin regions were tested. A sliding window of 50 � 50 pixels ismoved across the image region by 10 pixels overlapping in each trans-location. In each sliding position, the 50 � 50 window is assigned by theSVM model as belonging to the AK class or the healthy skin class. TheSVM model has been previously trained by regions from all patientsexcept the patient under testing. The skin region under testing was finallyrecognized as AK or healthy skin according to the majority class assignedto the 50 � 50 windows of that region.

To visualize the results, SVM outputs were converted to posteriorprobabilities in the range [0 1] [28]. In each sliding position a pseudo-color was assigned using a color map that corresponds a probability valueof 0 to blue (healthy class) and a probability value of 1 to yellow (AK

Fig. 2. ROI selection for model implementation starting from areas demarcated by experts in a cdelimited by black and healthy regions delimited by blue borders. A round color coding label (rpixels per mm. (b) From demarcated skin areas standardized 50 � 50 pixel ROIs were extractclassifier. The greater rectangular field delimited by the four black dots corresponds to a 25 cmreader is referred to the web version of this article.)

52

class). Explanatory examples are given in Fig. 3.SVM models were implemented using Matlab (R2015b, The Math-

Works) and run on a personal computer with a 64-bit operating system,2.4 GHz CPU and 12 GB memory.

2.3. Color texture analysis

The clinical diagnosis of AK relies on the recognition of certainmacromorphological characteristics. A typical AK appears as a white,scaly lesion of variable thickness with surrounding redness [1,3]. Uponpalpation AK are detected due to their distinctive sandpaper-like texture,supporting our strategy of using color texture analysis for the discrimi-nation of facial AK from normal skin.

Two main approaches to feature extraction for color texture analysisare often used and consist of:

linical photograph. (a) Patient photograph with regions demarcated by experts: AK lesionsed) was used as the internal standard – all images were rescaled to an equal final size of 50ed and (c) from each ROI color LBP and texton features were extracted to feed the SVM2 treatment area. (For interpretation of the references to colour in this figure legend, the

Table 1Number of AK and healthy ROIs (50 � 50 pixels) extracted from each patient.

Patient AK ROIs Normal skin ROIs

1 147 2162 425 12753 87 3404 230 8805 250 3756 546 18357 225 608 537 1939 66 19310 536 33211 197 4712 698 121713 459 166014 87 56315 12 1816 330 65817 194 135918 72 10819 586 159520 168 68321 66 4822 92 260

Total 6010 13915

Fig. 3. Examples of application of the discrimination model to visualize and quantify AKburden in selected skin regions. Top: (a) A tested AK skin region of 250 � 200 pixels (leftside) was separated into 336 50 � 50 pixel windows (10 pixel overlap of adjacent win-dows). (b) In each window SVM posterior probability was color-mapped to better visualizethe automatic skin recognition results. The majority of the windows were positives (309/336; 92%) and the whole skin region was recognized as an AK lesion. Bottom: (a) A testedhealthy skin region of 200 � 200 pixels was separated into 256 50 � 50 pixel windows (10pixel overlap). (b) Color-mapped SVM posterior probabilities. The majority of the win-dows were negatives (231/256; 90.2%) and the whole skin region was recognized ashealthy. (For interpretation of the references to colour in this figure legend, the reader isreferred to the web version of this article.)

P. Spyridonos et al. Computers in Biology and Medicine 88 (2017) 50–59

i. Applying gray-level texture analysis techniques to each color channelseparately

ii. Extracting textural information from the luminance plane along withpure chrominance features.

Both methods have resulted in substantial improvement in the per-formance of standard gray-level texture analysis techniques [29].

In the present study, we adopted the first method of extending thegray-level texture to channel-wise color texture analysis. More specif-ically, we considered channel-wise color LBP and raw pixel representa-tion textons. Both LBP and texton-based approaches have exhibitedexcellent texture classification performance on standard texture data-bases [30,31].

RGB color space is the standard format for displaying color digitalimages. A problem with this color representation is that it is device-dependent. Moreover, RGB values are very sensitive to illumination in-tensity and color changes because of the high correlation among the R, Gand B components. Thus, although RGB is suitable for color display, it isnot preferred for color analysis. RGB space can achieve satisfactory colordiscrimination accuracy under controlled illumination, however, this isnot retained under variable illumination conditions. Additionally, AKappear with variable colors. Thus absolute color evaluation is not usefulin image analysis. For improved color processing RGB is commonlytransformed to alternative spaces [29,32–34]. In this paper, along withRGB, YIQ and CIE-Lab color spaces were also investigated.

In the YIQ color model, the Y component is a measure of colorluminance. The I component represents the hue and the Q componentrepresents the saturation of the image. The YIQ space can partly reducethe correlation of the red, green and blue components in an image.

In the CIE-Lab color space, the L defines the lightness or the intensityof a color whereas a and b are the chromaticity components. Lab spacecan also control color and intensity information more independently.

YIQ has been proved to be an effective color space transformation forfacial image analysis applications whereas CIE-Lab color space associateswell with human vision systems [33].

2.4. Review of LBP features

Assuming a local circular region of a radius R consisting of P neigh-bors, the LBP operator is estimated by taking the difference of the centerpixel with respect to its neighbors:

53

LBPP;R ¼XP�1

p¼0

s�gp � gc

�2p (4)

s�gp � gc

� ¼�1 gp � gc0 gp < gc

(5)

where gc and gp denote the values of the central pixel and its neighbor,respectively, and p is the index of the neighbor. The values of neighborsthat are not in the center of grids are estimated by interpolation.

The U value of an LBP pattern is defined as the number of spatialtransitions (bitwise 0/1 changes):

UðLBPP;RÞ ¼ jsðgP�1 � gcÞ � sðg0 � gcÞj þXP�1

p¼1

��s�gp � gc�� s

�gp�1

� gc���

(6)

It has been shown that only “uniform” patterns (U � 2) are essentialpatterns of local image texture [35]. For example, bit strings 00000000(0 transitions) and 01111000 (two transitions) are uniform while10001001 (four transitions) and 01010110 (six transitions) are not. Non-uniform patterns (U>2) are grouped under the “miscellaneous” label.The mapping from LBPP;R to LBPu2

P;R reduces the different output valuesfrom 2p to P� ðP� 1Þþ 3 (superscript u2 means that the uniform pat-terns have U � 2).

Further, to extract rotation-invariant LBP descriptors, the locallyrotation-invariant pattern was adopted. This is defined as follows [36]:

P. Spyridonos et al. Computers in Biology and Medicine 88 (2017) 50–59

LBPriu2P;R ¼

8<:

PP�1

p¼0s�gp � gc

�if UðLBPP;RÞ � 2

Pþ 1 otherwise(7)

The mapping from LBPP;R to LPBriu2P;R results in P þ 2 different

output values.After estimating the LBPriu2P;R pattern in each pixel (i, j), the texture of

an ROI of 50 � 50 pixels is represented by building a histogram:

HðkÞ ¼X50i¼1

X50j¼1

f�LBPriu2

P;R ði; jÞ; k�; kε ½0;K� (8)

f ðx; yÞ ¼�1; x ¼ y0; otherwise

(9)

The aforementioned LBP operations (Eqs. (7)–(9)) are applied to eachcomponent of the YIQ (or CIE-Lab) color space. The final color LBP his-togram of each ROI is obtained by concatenating the d individual channelLBP histograms into a single column vector with d� ðPþ 2Þ elements,where d is the number of color channels used from the YIQ (or CIE-Lab)color space.

2.5. Review of textons using image patch exemplars

There are two main approaches proposed in the literature for texton-based texture description, filter banks [37] or raw pixel representation[31]. Raw pixel representation has recently gained ground, exhibitingsimilar or even superior classification results compared with multi-scale,multi-orientation filter banks [31,38].

Irrespective of the approach used to describe texture information,implementing textons consists of the following processing steps:initially, a codebook (dictionary) of textons is created using a clus-tering algorithm such as k-means. To construct the texton codebookusing the raw pixel representation approach, small-sized N � N(3 � 3 in the present case) local patches are extracted from each ROI inthe training set and the raw pixel intensities of these square neigh-borhoods are column-reordered to form a d� N2 dimensional vectorwhere d is the number of color channels used from the YIQ (or CIE-Lab) color space. All the patches from the training ROIs are aggre-gated and clustered. The set of cluster centers comprises the textondictionary. Next, a texton histogram is computed for each ROI in thetraining set. For this, small patches (3 � 3) are extracted by sliding awindow over each training ROI. Then, a histogram of textons iscomputed by comparing every patch representation in that ROI with

Fig. 4. Overview of texton-based feature extraction and classificat

54

all textons in the dictionary using the Euclidean distance. A schematicoverview of the herein applied texton feature extraction and classifi-cation procedures is given in Fig. 4.

2.6. Treatment evaluation of AK

Photographs at baseline and during follow-up were acquired fromeight patients. Treatment approaches were individually adapted ac-cording to good clinical practice and patients' needs. For thin or slightlyhyperkeratotic AK either of two modalities was employed: ingenolmebutate 0.015% gel (Picato®, Leo) once daily for 3 consecutive days totreat 25 cm2 skin areas with AK or a modified daylight PDT method for“field” treatment (2 h closed application of methyl aminolevulinate 16%cream [Metvix®, Galderma] followed by 30 min exposure to daylight)[39,40]. Cryosurgery (open spray, liquid N2, two cycles of 10 s each) wasapplied in solitary, hyperkeratotic AK, or in combination with topicalapplication of imiquimod 5% cream (immunocryosurgery) when fieldtreatment was indicated [41].

During image acquisition attention was given to capture almost thesame field before and after treatment. For AK quantification, an internalpoint was selected manually as a reference point in the baseline image,and the examined region was evaluated by the discrimination model.The polar coordinates of the upper left corner of the region to thereference point were estimated (Fig. 5, top). In the follow-up image, thesame internal point was manually selected, and the corresponding re-gion was autoselected based on the same polar coordinates used in theprevious baseline image for the upper left corner of the region box(Fig. 5, bottom).

3. Results

3.1. ROI-level evaluation of the SVMLBP model

Table 2 summarizes the performance of the SVMLBP model in threecolor spaces, namely RGB, YIQ and CIE-Lab. Employing the leave-one-patient-out method, YIQ and CIE-Lab color spaces exhibited similardiscrimination accuracy, and generally both outperformed RGB. To selectthe most informative color components of a color space, we tested allpossible color combinations within each space in one, two and three-dimensional configurations. Results for YIQ and CIE-Lab spaces aregiven in Table 2.2 and 2.3. The combination of Y and I components fromthe YIQ color space yielded the highest discrimination rates with sensi-tivity and specificity of 80.2% and 80.1%, respectively (Table 2.2).

The best configuration of the color texture operator LBPriu2P;R was

achieved with circular neighborhood P ¼ 8 and radius R ¼ 1.

ion. Dashed paths signify the classifier evaluation procedure.

Fig. 5. Application of the model to quantify treatment effects (example from patient 1, Table 4). Top: Finding at baseline. (a) A tested region box of 450 � 350 pixels was selected and thepolar coordinates of the upper left corner were estimated using an internal reference point (green point). (b) The selected region was separated into 1271 50 � 50 pixel windows (10 pixeloverlap) and was model-evaluated and quantified; 43.2% (549/1271) of the windows are positive. Bottom: The same area at a follow-up appointment 4 months post-treatment. (a) Thetested region box was re-identified and autoselected (red points) using the polar coordinates of the upper left corner to the reference point (green point). (b) The region was model-evaluated and quantified; only 9.8% of the windows (25/1271) were now positive, corresponding to a 77.3% improvement in AK burden. (For interpretation of the references tocolour in this figure legend, the reader is referred to the web version of this article.)

P. Spyridonos et al. Computers in Biology and Medicine 88 (2017) 50–59

Finally, the SVMLBP classification model was implemented using anRBF kernel parameter γ equal to 10 and box constraint parameter C equalto 1.

Since the number of healthy skin patterns was selected equal to that ofAK patterns, the prior probabilities and classification costs were set equalfor both classes.

3.2. ROI-level evaluation of the SVMTextons model

To construct the codebook of textons, small local patches of 3 � 3pixels were randomly extracted from each ROI in the training set. Each3 � 3 color patch forms a vector in a d � 9 dimensional feature spacewhere d is the number of color channels used from the YIQ color model.These patch vectors were clustered using the k-means algorithm withk ¼ 35 codewords.

Comparative recognition results of AK lesions using texton de-scriptors are given in Table 2.4. The best accuracy levels were achievedusing all components of the YIQ color space, reaching 79.6% and 73.8%for AK and healthy skin, correspondingly.

3.3. Region-level evaluation of the SVMLBP model

The SVMLBPriu28;1discrimination model, using as inputs concatenated

individual Y and I color channel LBP histograms from 50 � 50 slidingwindows, was tested for correctly recognizing skin regions from eachpatient; 141 out of 157 AK regions (89.8%) and 198 out of 216 healthyskin regions (91.7%) were correctly recognized. Results are summarizedin Table 3. Representative examples of misrecognized skin regions aredepicted in Fig. 6.

55

3.4. Evaluation of AK treatment outcome: a preliminary study

Photographs at baseline and during follow-up were acquired fromeight patients who were not included in the training set of themodel (Table 4).

In six out of the eight patients, multiple lesion fields were evaluatedresulting in a total of 32 skin areas selected for the evaluation of treat-ment outcomes. Fig. 5 depicts the evaluation outcome of selected skinregions, before and after treatment (patient 1: multiple lesions; patient 2:a single lesion). Based on the classifier results, individualized, bestpractice, minimally invasive interventions yielded an average AK burdenreduction of 83.6% (Table 4).

4. Discussion

Clinical photography is widely available and is currently consideredessential for the documentation of dermatologic conditions. In conjunc-tion with computer vision techniques it could be upgraded to a funda-mental and cost-effective tool for the evaluation of various skin lesions,monitoring disease progression as well as quantifying treatment in-terventions [22,23,42–47].

In the present study, we explored the utility of clinical photographs inthe assessment of AK, a frequent cutaneous premalignancy. We employedan SVM classifier and color texture features to discriminate between AKlesions and healthy skin and to quantify field-directed treatmentinterventions.

In a different, computer-aided diagnosis approach, Boone et al. [16]have quantified OCT images of facial AK/sSCC lesions. Applying receiveroperating curves to determine the cut-off values of the optical charac-teristics of lesions, they proposed a decision tree diagnostic algorithm for

Table 2SVMLBP and SVMTextons model performance in color spaces. 2.1 Classification rates usingcolor operator LBPriu2

8;1 in different color spaces. 2.2 Classification results using coloroperator LBPriu2

8;1 and color channel combinations in the YIQ color space. 2.3 Classification

results using color operator LBPriu28;1 and color channel combinations in CIE-Lab color. 2.4

Classification results using color texton operator and color channel combinations in theCIE-Lab color space. The approach with the best discriminating performance is depicted inbold.

Sensitivity (%) Specificity (%)

2.1 LBPriu28;1

RGB 75.5 78.3YIQ 79 82.3CIE-Lab 78 82

2.2 LBPriu28;1

YIQ-Y 79 75.8YIQ-I 73 66.7YIQ-Q 68.5 68.40YIQ-IQ 71.7 75.3YIQ-YI 80.2 80.1YIQ-YQ 75.8 79

2.3 LBPriu28;1

LAB-L 79.3 76.1LAB-A 76.3 73.20LAB-B 73.2 65.6LAB-AB 78 79LAB-LA 78.3 80.3LAB-LB 77.3 78.7

2.4 Textons

YIQ-Y 68.6 79.2YIQ-I 63.7 72.1YIQ-Q 73.3 67.2YIQ-IQ 73 71.4YIQ-YI 69.7 76.9YIQ-YQ 70.5 73.3YIQ-YIQ 79.6 73.8

Table 3Confusion matrix for АΚ versus healthy skin regions; 141/157 AK regions (89.8%) and198/216 healthy regions (91.7%) were correctly recognized.

Clinically diagnosed regions Model-recognized regions Total

AK Healthy skin

AK 141 (89.8%) 16 157Healthy skin 18 198 (91.7%) 216

P. Spyridonos et al. Computers in Biology and Medicine 88 (2017) 50–59

the discrimination of sun-damaged facial skin from AK subtypes andsSCC. The diagnostic accuracy was high (>90%), yet these optimumoutcomes were acquired only in the training set. Horn et al. [17] havestudied RCM criteria for the discrimination of AK from normal skin. Twoobservers visually assessed the presence or absence of selected RCMfeatures and classification trees were used to differentiate AK fromnormal skin. Thus with a single RCM feature (“irregular keratinocyte cellborders”) sensitivity and specificity of 86.67% and 85%, respectively, inthe discrimination of AK from healthy skin were achieved. With theintegration of additional features into the model, specificity increased upto 90%. However, when considered separately, none of the evaluatedRCM criteria were more than moderately reliable for the recognition ofAK (interobserver agreement; kappa statistic). Huerta-Brogeras et al.[18] have proposed a diagnostic algorithm that combines the dermo-scopic signs “follicular openings” and “erythematous pseudo network”for the diagnosis of AK with 95.6% sensitivity and 95% specificitycompared to histopathology. However, evaluation of dermoscopic im-ages was subjective, based exclusively on clinical assessments by experts,and interobserver variability was not reported. McDonagh et al. [19]have used a dimensional imaging dense stereo capture system to obtain3D shape features to distinguish between neoplastic non-melanoma skinlesions automatically. Feeding depth and color image features into aBayesian classifier, they demonstrated improved classification rates forfive common classes of skin lesions: AK, sSCC, BCC, seborrhoeic keratosisand melanocytic nevus. They reported an overall method accuracy of83%. Regarding the AK class, the recognition rate was 100% with,however, the restriction that only 11 samples were analyzed.

The application of standard clinical photography in the evaluation ofkeratinocytic premalignancies has been previously investigated by Bal-lerini et al. [22]. A K-NN hierarchical scheme, using color and textureinformation from clinical photographs, attained an overall classification

56

accuracy of 74% over the five most common benign and malignant skingrowths (malignant melanoma, AK, BCC, melanocytic nevus/mole, sSCCand seborrhoeic keratosis). Regarding AK, most test images were mis-classified as BCC. In a more recent study, Hames et al. [23] proposed amethod for automatic AK border delineation on clinical photographsusing color space transformations and morphological operations. Auto-matic segmentation was compared with expert annotation of the pic-tures, also taking into account intraobserver variability. Correlationbetween automatic and manual AK identification was 0.62 for face and0.51 for arm lesions. The sensitivity of automatic detection was 39.5%and 53.1% and the positive predictive values 13.9% and 39.8% for facialand arm AK, respectively.

AK arise multifocally in the initiated epidermis of photo-damagedskin as small clusters of mutated cells [48] that may expand horizontal-ly and vertically within the epidermis for various periods until theybecome clinically evident [49]. Consequently, the affected skin harborsmultiple subclinical disease foci, in addition to clinically recognizable AK[50]. Thus, in most cases, the presence of overt AK in a skin area indicatesthe presence of an underlying skin cancerization field [51]. “Field ther-apies” such as PDT and topical ingenol mebutate are principally designedto target exactly these skin areas as a whole, addressing obvious as well assubclinical premalignant lesions [52]. For these reasons, in the presentwork, we did not attempt to strictly define the borders of individual,clinically demarcated AK lesions but rather addressed the discriminationof AK-affected areas from healthy skin and quantification of the cumu-lative AK burden within selected skin areas.

The proposed SVM classification model employing LBP color texturefeatures exhibited sensitivity and specificity over 80% at both ROI- andregion-based levels (Tables 2.2 and 3) and was also validated in eightpatients not included in the training set (Table 4, Fig. 7). In this lattergroup the model adequately enabled the quantification of treatment ef-ficacy and recorded an averaged 83.6% AK reduction.

Commenting misclassification, the primary source of false negativeswere lesions with subtle texture sub-areas interposed by smoothed areaswith erythema (Fig. 6, top). However, this “pattern” of lesions appearedonly in one patient (out of 22), and possibly our model was insufficientlytrained to recognize these lesions. Following, the majority of false posi-tives (Fig. 6, bottom) was mainly recorded from patients with severephotodamage and field carcinogenesis with multiple concurrent kerati-nocytic carcinomas. Though, this clinical background also challenges thecapability of experts to delineate lesions from healthy areas.

Herein, color LBP descriptors have been employed for the identifi-cation of morphological alterations of AK lesions. LBPriu2

8;1 color texturefeatures extracted from the Y and I components of the YIQ color spaceresulted in the best feature vector to use in the discrimination model.Color LBP and texton descriptors outperformed the conventional gray-scale texture analysis (Tables 2.2 and 2.4). Notably, the luminancecomponent Y of the YIQ color space represents grayscale information.

Our results are in line with the current findings indicating that colorinformation plays an important role in texture analysis and can be used toimprove recognition performance considerably [29,53]. Compared totextons, the LBP approach exhibited better identification performance inaddition to shorter computation times. A possible drawback of thisapproach is that patch exemplars are not rotationally invariant. However,it has been demonstrated that incorporating rotation-invariant textons,improves, even slightly, their classification performance [31]. Further-more, LBP descriptors have distinctive robustness against monotonic

Fig. 6. Examples of misrecognized skin regions. Top: (a) A tested AK skin region of 220 � 222 pixels was separated into 256 50 � 50 pixel windows (10 pixel overlap). (b) The majority ofthe windows were negatively misclassified (161/256; 62.9%), and the whole skin region was misrecognized as healthy skin (False negative). Bottom: (a) A tested healthy skin region of316 � 368 pixels was separated into 806 50 � 50 pixel windows (10 pixel overlap). (b) The majority of the windows were positively misclassified (687/806; 85.2%), and the whole regionwas misrecognized as AK lesion (False positive).

Table 4Quantitative evaluation of AK lesions before and after treatment using the SVMLBP discrimination model.

Patient Sex Age Treatment AK localization Follow-up (months) AK burden in %a

Before treatment After treatment AK reduction Mean AKreduction

1 M 85 Daylight PDTb, imiquimod,cryosurgery

Left mandible-preauricular area 4 43.2 9.8 77.3 71.857.2 16.9 70.524.2 7.8 67.8

Forehead (left half) 4 72.9 37.1 49.1 74.643.3 0 100.0

Forehead (right half) 7 43.4 0 100.0 95.966.7 5.5 91.8

2 M 75 Cryosurgery Right cheek 10 95.7 21.5 77.5 77.53 M 79 Ingenol mebutate Left cheek 2 96 28 70.8 82.2

71 6.5 90.893 14 84.9

Right cheek 2 64 32 50.0 74.263 1 98.4

Forehead 2 61 0 100.0 99.0100 2 98.0

4 M 82 Imiquimod, cryosurgery Right cheek 4 45.2 5.5 87.8 87.85 M 79 Ingenol mebutate Occipital 12 49.4 23.2 53.0 82.5

56 2.7 95.240.6 13.5 66.766.6 1.5 97.772.2 0 100.0

6 M 78 Ingenol mebutate Right cheek 2 98 20 79.6 73.695.5 44 53.992 11.7 87.3

7 M 89 Cryosurgery, ingenol mebutate Forehead (left half) 2 18.7 2 89.3 92.529.6 2.8 90.562.5 10.7 82.968.3 0 100.064.5 0 100.0

8 M 84 Cryosurgery, ingenol mebutate Temporoparietal region (left) 2 66.7 7.1 89.4 89.4Occipital 2 66 11.6 82.4 85.9

63.4 6.7 89.4Mean 64.1 10.8 83.5 83.6

a % skin area recognized as AK.b Photodynamic therapy.

P. Spyridonos et al. Computers in Biology and Medicine 88 (2017) 50–59

57

Fig. 7. Monitoring skin regions before (a) and after treatment (b). The first five examplesfrom the top were from patient 1 and the last ones from patient 2 (Table 4). In each pair offigures, the left one is the selected region of the clinical photograph and the right onedepicts the AK burden in this field. (c) Employed color bar: a pseudocolor was assignedusing a color map that corresponds the probability value of 0 (healthy class) to blue andthe probability value of 1 to yellow (AK class). (For interpretation of the references tocolour in this figure legend, the reader is referred to the web version of this article.)

P. Spyridonos et al. Computers in Biology and Medicine 88 (2017) 50–59

illumination changes and this applies in clinical practice where imagesare captured under varying illumination conditions. However, byassuming that the illumination in a small ROI of 50 � 50 pixels isapproximately uniform (monotonic), we can suggest that LBP descriptorscould accommodate potential local intensity fluctuations.

Regarding time efficiency, the computation time of an LBP histogramin a 50 � 50 pixel ROI was 0.02 s, whereas the corresponding time of anequivalent texton histogram was 0.08 s. To give a concrete examplecomparing the time efficiency of LBP- and texton-based feature extrac-tion we assumed the assessment of a skin region of 250 � 200 pixelswhich is a skin area of 50 � 40 mm2 (e.g. Fig. 3, top). This region wasseparated into 336 50 � 50 pixel windows. Consequently, the estimationtime of the AK burden for this example was about 6 s using LBP and about36 s using the texton approach.

In conclusion, clinical photography combined with relevant imageanalysis algorithms is a promising non-invasive, cost-effective moni-toring tool for the evaluation of field-directed treatment interventions forAK. Moreover, the discrimination power and the fast computation ofcolor LBP features render the herein proposed SVMLBP model a viable andreal-time method for evaluating treatment outcomes.

Appendix A. Supplementary data

Supplementary data related to this article can be found at http://dx.doi.org/10.1016/j.compbiomed.2017.07.001.

References

[1] L.H. Goldberg, A.J. Mamelak, Review ofactinic keratosis. Part I: etiology,epidemiology and clinical presentation, J. Drugs Dermatol. 9 (2010) 1125–1132.

[2] S.J. Salasche, Epidemiology of actinic keratoses and squamous cell carcinoma,J. Am. Acad. Dermatol 42 (2000) 4–7.

[3] J. Malvehy, A new vision of actinic keratosis beyond visible clinical lesions, J. Eur.Acad. Dermatol. Venereol. 29 (2015) 3–8, http://dx.doi.org/10.1111/jdv.12833.

[4] V.D. Criscione, M.A. Weinstock, M.F. Naylor, C. Luque, M.J. Eide, S.F. Bingham,Department of veteran affairs topical tretinoin chemoprevention trial group, actinickeratoses: natural history and risk of malignant transformation in the veteransaffairs topical tretinoin chemoprevention trial, Cancer 115 (2009) 2523–2530,http://dx.doi.org/10.1002/cncr.24284.

58

[5] P.J.F. Quaedvlieg, E. Tirsi, M.R.T.M. Thissen, G.A. Krekels, Actinic keratosis: howto differentiate the good from the bad ones? Eur. J. Dermatol. 16 (2006)335–339.

[6] D.S. Rigel, L.F. Stein Gold, The importance of early diagnosis and treatment ofactinic keratosis, J. Am. Acad. Dermatol. 68 (2013) S20–S27, http://dx.doi.org/10.1016/j.jaad.2012.10.001.

[7] R. Marks, The role of treatment of actinic keratoses in the prevention of morbidityand mortality due to squamous cell carcinoma, Arch. Dermatol. 127 (1991)1031–1033.

[8] R.M. Hurwitz, L.E. Monger, Solar keratosis: an evolving squamous cell carcinoma.Benign or malignant? Dermatol. Surg. 21 (1995) 184, http://dx.doi.org/10.1111/j.1524-4725.1995.tb00141.x.

[9] K.C. Lee, R. Lew, M.A. Weinstock, Improvement in precision of counting actinickeratoses, Br. J. Dermatol. 170 (2014) 188–191, http://dx.doi.org/10.1111/bjd.12629.

[10] S.C. Chen, N.D. Hill, E. Veledar, S.M. Swetter, M.A. Weinstock, Reliability ofquantification measures of actinic keratosis, Br. J. Dermatol. 169 (2013)1219–1222, http://dx.doi.org/10.1111/bjd.12591.

[11] M. Ianhez, L.F.F. Fleury Junior, E. Bagatin, H.A. Miot, The reliability of countingactinic keratosis, Arch. Dermatol. Res. 305 (2013) 841–844, http://dx.doi.org/10.1007/s00403-013-1413-y.

[12] M. Emre Celebi, W.V. Stoecker, R.H. Moss, Advances in skin cancer image analysis,Comput. Med. Imaging Graph 35 (2011) 83–84, http://dx.doi.org/10.1016/j.compmedimag.2010.11.005.

[13] I. Maglogiannis, C.N. Doukas, Overview of advanced computer vision systems forskin lesions characterization, IEEE Trans. Inf. Technol. Biomed. 13 (2009) 721–733,http://dx.doi.org/10.1109/TITB.2009.2017529.

[14] K. Korotkov, R. Garcia, Computerized analysis of pigmented skin lesions: a review,Artif. Intell. Med. 56 (2012) 69–90, http://dx.doi.org/10.1016/j.artmed.2012.08.002.

[15] T.M. Jorgensen, A. Tycho, M. Mogensen, P. Bjerring, G.B.E. Jemec, Machine-learning classification of non-melanoma skin cancers from image features obtainedby optical coherence tomography, Ski. Res. Technol. 14 (2008) 364–369, http://dx.doi.org/10.1111/j.1600-0846.2008.00304.x.

[16] M.A.L.M. Boone, M. Suppa, A. Marneffe, M. Miyamoto, G.B.E. Jemec, V. DelMarmol, A new algorithm for the discrimination of actinic keratosis fromnormal skin and squamous cell carcinoma based on in vivo analysis of opticalproperties by high-definition optical coherence tomography, J. Eur. Acad.Dermatol. Venereol. 30 (2016) 1714–1725, http://dx.doi.org/10.1111/jdv.13720.

[17] M. Horn, A. Gerger, V. Ahlgrimm-Siess, W. Weger, S. Koller, H. Kerl, H. Samonigg,J. Smolle, R. Hofmann-Wellenhof, Discrimination of actinic keratoses from normalskin with reflectance mode confocal microscopy, Dermatol. Surg. 34 (2008)620–625, http://dx.doi.org/10.1111/j.1524-4725.2008.34195.x.

[18] M. Huerta-Brogeras, O. Olmos, J. Borbujo, A. Hern�andez-Nú~nez, E. Casta~no,A. Romero-Mat�e, D. Martínez-S�anchez, C. Martínez-Mor�an, Validation ofdermoscopy as a real-time noninvasive diagnostic imaging technique for actinickeratosis, Arch. Dermatol. 148 (2012) 1159, http://dx.doi.org/10.1001/archdermatol.2012.1060.

[19] S. McDonagh, R. Fisher, J. Rees, Using 3D information for classification of non-melanoma skin lesions, in: Proc. Med. Image Underst. Anal, BMVA Press, 2008,pp. 164–168. BT–Proc. Medical Image Understanding and.

[20] B. Jung, Dermatological feasibility of multimodal facial color imaging modality forcross-evaluation of facial actinic keratosis, Ski. Res. Technol. (2011) 4–10, http://dx.doi.org/10.1111/j.1600-0846.2010.00464.x.

[21] C. Wassef, B.K. Rao, Uses of non-invasive imaging in the diagnosis of skin cancer: anoverview of the currently available modalities, Int. J. Dermatol. 52 (2013)1481–1489, http://dx.doi.org/10.1111/ijd.12159.

[22] L. Ballerini, R. Fisher, B. Aldridge, J. Rees, A color and texture based hierarchical K-NN approach to the classification of non-melanoma skin lesions, Color Med. ImageAnal. (2013) 63–86, http://dx.doi.org/10.1007/978-94-007-5389-1_4.

[23] S.C. Hames, S. Sinnya, J.M. Tan, C. Morze, A. Sahebian, H.P. Soyer, T.W. Prow,Automated detection of actinic keratoses in clinical photographs, PLoS One 10(2015) 1–12, http://dx.doi.org/10.1371/journal.pone.0112447.

[24] H. Byun, S.-W. Lee, Applications of support vector machines for pattern recognition:a survey, LNCS 2388 (2002) 213–236.

[25] J.M. Moguerza, A. Mu~noz, Support vector machines with applications, Stat. Sci. 21(2006) 322–336, http://dx.doi.org/10.1214/088342306000000493.

[26] A. Masood, A.A. Al-Jumaily, Computer aided diagnostic support system for skincancer: a review of techniques and algorithms, Int. J. Biomed. Imaging 2013 (2013)323268, http://dx.doi.org/10.1155/2013/323268.

[27] C. Cortes, V. Vapnik, Support-vector networks, Mach. Learn 20 (1995) 273–297,http://dx.doi.org/10.1007/BF00994018.

[28] J.C. Platt, J.C. Platt, Probabilistic outputs for support vector machines andcomparisons to regularized likelihood methods, Adv. Large Margin Classif. (1999)61–74.

[29] A. Drimbarean, P.F. Whelan, Experiments in colour texture analysis, PatternRecognit. Lett. 22 (2001) 1161–1167, http://dx.doi.org/10.1016/S0167-8655(01)00058-7.

[30] T. Ojala, M. Pietikainen, T. Maenpaa, Multiresolution gray-scale and rotationinvariant texture classification with local binary patterns, IEEE Trans. Pattern Anal.Mach. Intell. 24 (2002) 971–987, http://dx.doi.org/10.1109/TPAMI.2002.1017623.

[31] M. Varma, A. Zisserman, A statistical approach to material classification usingimage patch exemplars, IEEE Trans. Pattern Anal. Mach. Intell. 31 (2009)2032–2047, http://dx.doi.org/10.1109/TPAMI.2008.182.

P. Spyridonos et al. Computers in Biology and Medicine 88 (2017) 50–59

[32] P. Kakumanu, S. Makrogiannis, N. Bourbakis, A survey of skin-color modeling anddetection methods, Pattern Recognit. 40 (2007) 1106–1122, http://dx.doi.org/10.1016/j.patcog.2006.06.010.

[33] J. Yang, C. Liu, L. Zhang, Color space normalization: enhancing the discriminatingpower of color spaces for face recognition, Pattern Recognit. 43 (2010) 1454–1466,http://dx.doi.org/10.1016/j.patcog.2009.11.014.

[34] G. Kaur, K. Joshi, A brief survey about existed segmentation techniques inautomatic detection and segmentation of skin melanoma images, Int. J. Emerg. Res.Manag. Technol. (2015) 2278–9359.

[35] O. Lahdenoja, J. Poikonen, M. Laiho, Towards understanding the formation ofuniform local binary patterns, Int. Sch. Res. Not. 2013 (2013) e429347, http://dx.doi.org/10.1155/2013/429347.

[36] T. Ojala, M. Pietik€ainen, T. M€aenp€a€a, Gray scale and rotation invariant textureclassification with local binary patterns, In Pract. 1842, Computer Vision-ECCV(2000) 404–420, http://dx.doi.org/10.1007/3-540-45054-8.

[37] M. Varma, A. Zisserman, A statistical approach to texture classification from singleimages, Int. J. Comput. Vis. 62 (2005) 61–81, http://dx.doi.org/10.1007/s11263-005-4635-4.

[38] M.J. Gangeh, L. Sørensen, S.B. Shaker, M.S. Kamel, M. de Bruijne, Multiple classifiersystems in texton-based approach for the classification of CT images of lung, in:A.C. Bjoern Menze, Georg Langs,Zhowen Tu (Eds.), Med. Comput. Vision. Recognit.Tech. Appl. Med. Imaging Int. MICCAI Work. MCV 2010, Beijing, China, Sept. 20,2010, Revis. Sel. Pap, Springer Berlin Heidelberg, 2010, pp. 153–163, http://dx.doi.org/10.1007/978-3-642-18421-5_15.

[39] S.R. Wiegell, M. Haedersdal, P.A. Philipsen, P. Eriksen, C.D. Enk, H.C. Wulf,Continuous activation of PpIX by daylight is as effective as and less painful thanconventional photodynamic therapy for actinic keratoses; a randomized, controlled,single-blinded study, Br. J. Dermatol. 158 (2008) 740–746, http://dx.doi.org/10.1111/j.1365-2133.2008.08450.x.

[40] D.M. Rubel, L. Spelman, D.F. Murrell, J.-A. See, D. Hewitt, P. Foley, C. Bosc,D. Kerob, N. Kerrouche, H.C. Wulf, S. Shumack, Daylight photodynamic therapywith methyl aminolevulinate cream as a convenient, similarly effective, nearlypainless alternative to conventional photodynamic therapy in actinic keratosistreatment: a randomized controlled trial, Br. J. Dermatol. 171 (2014) 1164–1171,http://dx.doi.org/10.1111/bjd.13138.

[41] G. Gaitanis, I.D. Bassukas, Immunocryosurgery for nonmelanoma skin cancer:applications and practical tips, in: P. Pasquali (Ed.), Cryosurgery, Springer BerlinHeidelberg, Berlin, Heidelberg, 2015, pp. 245–258, http://dx.doi.org/10.1007/978-3-662-43939-5_20.

[42] P. Spyridonos, G. Gaitanis, M. Tzaphlidou, I.D. Bassukas, Spatial fuzzy c-meansalgorithm with adaptive fuzzy exponent selection for robust vermilion borderdetection in healthy and diseased lower lips, Comput. Methods Programs Biomed.114 (2014) 291–301, http://dx.doi.org/10.1016/j.cmpb.2014.02.017.

59

[43] P. Spyridonos, G. Gaitanis, I.D. Bassukas, M. Tzaphlidou, Evaluation of vermillionborder descriptors and relevance vector machines discrimination model formaking probabilistic predictions of solar cheilosis on digital lip photographs,Comput. Biol. Med. 63 (2015) 11–18, http://dx.doi.org/10.1016/j.compbiomed.2015.04.024.

[44] P. Spyridonos, G. Gaitanis, I.D. Bassukas, M. Tzaphlidou, Gray Hausdorff distancemeasure for medical image comparison in dermatology: evaluation of treatmenteffectiveness by image similarity, Skin. Res. Technol. 19 (2013) e498–506, http://dx.doi.org/10.1111/srt.12001.

[45] J. Fern�andez Alc�on, C. Ciuhu, W. ten Kate, A. Heinrich, N. Uzunbajakava,G. Krekels, D. Siem, G. de Haan, Automatic imaging system with decision supportfor inspection of pigmented skin lesions and melanoma diagnosis, IEEE J. Sel.Top. Signal Process 3 (2009) 14–25, http://dx.doi.org/10.1109/JSTSP.2008.2011156.

[46] P.G. Cavalcanti, J. Scharcanski, Automated prescreening of pigmented skin lesionsusing standard cameras, Comput. Med. Imaging Graph 35 (2011) 481–491, http://dx.doi.org/10.1016/j.compmedimag.2011.02.007.

[47] W.-Y. Chang, A. Huang, C.-Y. Yang, C.-H. Lee, Y.-C. Chen, T.-Y. Wu, G.-S. Chen,Computer-aided diagnosis of skin lesions using conventional digital photography: areliability and feasibility study, PLoS One 8 (2013) e76212, http://dx.doi.org/10.1371/journal.pone.0076212.

[48] R.J. Berg, H.J. van Kranen, H.G. Rebel, A. de Vries, W.A. van Vloten, C.F. Van Kreijl,J.C. van der Leun, F.R. de Gruijl, Early p53 alterations in mouse skin carcinogenesisby UVB radiation: immunohistochemical detection of mutant p53 protein in clustersof preneoplastic epidermal cells, Proc. Natl. Acad. Sci. U. S. A. 93 (1996) 274–278.

[49] R.M. Szeimies, L. Torezan, A. Niwa, N. Valente, P. Unger, E. Kohl, S. Schreml,P. Babilas, S. Karrer, C. Festa-Neto, Clinical, histopathological andimmunohistochemical assessment of human skin field cancerization before andafter photodynamic therapy, Br. J. Dermatol. 167 (2012) 150–159, http://dx.doi.org/10.1111/j.1365-2133.2012.10887.x.

[50] C.J. Cockerell, Histopathology of incipient intraepidermal squamous cell carcinoma(“actinic keratosis”), J. Am. Acad. Dermatol. 42 (2000) 11–17.

[51] J. Einspahr, D.S. Alberts, M. Aickin, K. Welch, P. Bozzo, T. Grogan, M. Nelson,Expression of p53 protein in actinic keratosis, adjacent, normal-appearing, and non-sun-exposed human skin, Cancer Epidemiol. Biomarkers Prev. 6 (1997) 583–587.

[52] D. de Berker, J.M. McGregor, M.F. Mohd Mustapa, L.S. Exton, B.R. Hughes, BritishAssociation of Dermatologists' guidelines for the care of patients with actinickeratosis 2017, Br. J. Dermatol. 176 (2017) 20–43, http://dx.doi.org/10.1111/bjd.15107.

[53] J.Y. Choi, K.N. Plataniotis, Y.M. Ro, Using colour local binary pattern features forface recognition, in: 2010 IEEE Int. Conf. Image Process, IEEE, 2010,pp. 4541–4544, http://dx.doi.org/10.1109/ICIP.2010.5653653.